Why sampling from a multinomial distribution?

[pyun-ram]

Hi @ariseff , thanks for sharing the awesome code repo! :)
Your code works well, but I cannot understand why we need to do sampling from a multinomial distribution?
Can we replace it with argmax or just with no-sampling?

overcoming-catastrophic/model.py

Line 54 in cab6d86

class_ind = tf.to_int32(tf.multinomial(tf.log(probs), 1)[0][0])